IQ2_S_R4 (#156)

* iq2_s_r4: Zen4

* Minor

* iq2_s_r4: NEON

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>

6 files changed, 330 insertions, 0 deletions

diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index 2534b461..6964f5e6 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -428,6 +428,15 @@ typedef struct {
 } block_iq2_s;
 static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
 
+typedef struct {
+    ggml_half d[4];
+    uint8_t qs[QK_K/2];
+    uint8_t qh[QK_K/8];
+    uint8_t signs[QK_K/2];
+    uint8_t scales[QK_K/8];
+} block_iq2_s_r4;
+static_assert(sizeof(block_iq2_s_r4) == 4*sizeof(block_iq2_s), "wrong iq2_s_r4 block size/padding");
+
 // (Almost) "true" 3-bit quantization.
 // Due to the need to use blocks as per ggml design, it ends up using
 // 3.0625 bpw because of the 16-bit scale for each block of 256.
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 1f56ec06..bf028c0c 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -15201,6 +15201,7 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
         case GGML_TYPE_IQ2_XXS_R4: break;
         case GGML_TYPE_IQ2_XS_R4: break;
         case GGML_TYPE_IQ3_XXS_R4: break;
+        case GGML_TYPE_IQ2_S_R4: break;
         case GGML_TYPE_Q4_0_R4: break;
         case GGML_TYPE_Q5_0_R4: break;
         case GGML_TYPE_Q6_0_R4: break;
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index 0c3be11c..2cece547 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -1096,6 +1096,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
         .row_meta_size            = 0,
     },
+    [GGML_TYPE_IQ2_S_R4] = {
+        .type_name                = "iq2_s_r4",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_s),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_s_r4,
+        .from_float               = quantize_row_iq2_s_r4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq2_s_r4_ref,
+        .vec_dot                  = vec_dot_iq2_s_r4_q8_k,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+        .row_meta_size            = 0,
+    },
     [GGML_TYPE_IQ1_S] = {
         .type_name                = "iq1_s",
         .blck_size                = QK_K,
@@ -4270,6 +4283,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ6_K:         wtype = GGML_TYPE_IQ6_K;    break;
         case GGML_FTYPE_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
         case GGML_FTYPE_MOSTLY_IQ2_S:         wtype = GGML_TYPE_IQ2_S;    break;
+        case GGML_FTYPE_MOSTLY_IQ2_S_R4:      wtype = GGML_TYPE_IQ2_S_R4; break;
         case GGML_FTYPE_MOSTLY_Q4_0_4_4:      wtype = GGML_TYPE_Q4_0_4_4; break;
         case GGML_FTYPE_MOSTLY_Q4_0_4_8:      wtype = GGML_TYPE_Q4_0_4_8; break;
         case GGML_FTYPE_MOSTLY_Q4_0_8_8:      wtype = GGML_TYPE_Q4_0_8_8; break;
@@ -10814,6 +10828,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ2_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -11277,6 +11292,7 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ2_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -11437,6 +11453,7 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ2_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -14643,6 +14660,7 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ2_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -15043,6 +15061,7 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ2_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -15337,6 +15356,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ2_S_R4:
         case GGML_TYPE_Q4_0_4_4:
         case GGML_TYPE_Q4_0_4_8:
         case GGML_TYPE_Q4_0_8_8:
@@ -15960,6 +15980,7 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ6_K:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ2_S_R4:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_Q8_K64:
         case GGML_TYPE_Q8_K16:
@@ -22702,6 +22723,7 @@ void ggml_quantize_init(enum ggml_type type) {
     switch (type) {
         case GGML_TYPE_IQ2_XXS_R4: iq2xs_init_impl(GGML_TYPE_IQ2_XXS); break;
         case GGML_TYPE_IQ2_XS_R4:  iq2xs_init_impl(GGML_TYPE_IQ2_XS);  break;
+        case GGML_TYPE_IQ2_S_R4:   iq2xs_init_impl(GGML_TYPE_IQ2_S);   break;
         case GGML_TYPE_IQ2_XXS:
         case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_IQ2_S:
@@ -22786,6 +22808,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ3_XXS_R4:result = quantize_iq3_xxs_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ3_S:   result = quantize_iq3_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ2_S:   result = quantize_iq2_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ2_S_R4:result = quantize_iq2_s_r4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_S:   result = quantize_iq1_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ1_BN:  result = quantize_iq1_bn (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp
index a21700a9..081ebb57 100644
--- a/ggml/src/iqk/iqk_mul_mat.cpp
+++ b/ggml/src/iqk/iqk_mul_mat.cpp
@@ -3426,6 +3426,128 @@ static void mul_mat_iq2_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data
 }
 
 template <int nrc_y>
+static void mul_mat_iq2_s_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+    GGML_ASSERT(nrc_x%4 == 0);
+    Q8<nrc_y, block_q8_K> q8(info);
+    int nbl = n / QK_K;
+#ifndef HAVE_FANCY_SIMD
+    auto smask = _mm256_set1_epi64x(0x8040201008040201);
+    auto sign_shuffle = _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202, 0x0101010101010101, 0x0000000000000000);
+    auto m4 = _mm256_set1_epi8(4);
+#endif
+    __m256  acc[nrc_y] = {};
+#ifdef HAVE_FANCY_SIMD
+    __m256i shuffles[2] = {
+        _mm256_set_epi64x(0x0706070607060706, 0x0302030203020302, 0x0504050405040504, 0x0100010001000100),
+        _mm256_set_epi64x(0x0f0e0f0e0f0e0f0e, 0x0b0a0b0a0b0a0b0a, 0x0d0c0d0c0d0c0d0c, 0x0908090809080908)
+    };
+    __m256i isum[2*nrc_y] = {};
+#else
+    __m256i shuffles[4] = {
+        MM256_SET_M128I(_mm_set1_epi16(0x0302), _mm_set1_epi16(0x0100)),
+        MM256_SET_M128I(_mm_set1_epi16(0x0706), _mm_set1_epi16(0x0504)),
+        MM256_SET_M128I(_mm_set1_epi16(0x0b0a), _mm_set1_epi16(0x0908)),
+        MM256_SET_M128I(_mm_set1_epi16(0x0f0e), _mm_set1_epi16(0x0d0c)),
+    };
+    __m256i isum[nrc_y == 1 ? 4 : nrc_y] = {};
+#endif
+    __m256i qx[4];
+    auto grid = iq2s_grid;
+    for (int ix = 0; ix < nrc_x; ix += 4) {
+        auto iq2 = (const block_iq2_s_r4 *)((const char *)vx + (ix+0)*bx);
+        for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
+            auto dl = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq2[ibl].d));
+            auto d4 = _mm256_set_m128(dl, dl);
+            auto s32 = (const uint32_t *)iq2[ibl].scales;
+            auto ql = iq2[ibl].qs;
+            auto qh = iq2[ibl].qh;
+            for (int ib = 0; ib < QK_K/32; ++ib) {
+                qx[0] = _mm256_set_epi64x(grid[ql[ 3] | ((qh[0] << 2) & 0x300)], grid[ql[ 2] | ((qh[0] << 4) & 0x300)], grid[ql[ 1] | ((qh[0] << 6) & 0x300)], grid[ql[ 0] | ((qh[0] << 8) & 0x300)]);
+                qx[1] = _mm256_set_epi64x(grid[ql[ 7] | ((qh[1] << 2) & 0x300)], grid[ql[ 6] | ((qh[1] << 4) & 0x300)], grid[ql[ 5] | ((qh[1] << 6) & 0x300)], grid[ql[ 4] | ((qh[1] << 8) & 0x300)]);
+                qx[2] = _mm256_set_epi64x(grid[ql[11] | ((qh[2] << 2) & 0x300)], grid[ql[10] | ((qh[2] << 4) & 0x300)], grid[ql[ 9] | ((qh[2] << 6) & 0x300)], grid[ql[ 8] | ((qh[2] << 8) & 0x300)]);
+                qx[3] = _mm256_set_epi64x(grid[ql[15] | ((qh[3] << 2) & 0x300)], grid[ql[14] | ((qh[3] << 4) & 0x300)], grid[ql[13] | ((qh[3] << 6) & 0x300)], grid[ql[12] | ((qh[3] << 8) & 0x300)]);
+                ql += 16; qh += 4;
+                auto signs128 = _mm_loadu_si128((const __m128i*)iq2[ibl].signs + ib);
+                auto scales = _mm_set1_epi32(s32[ib]);
+                scales = _mm_and_si128(_mm_unpacklo_epi8(scales, _mm_srli_epi16(scales, 4)), _mm_set1_epi8(0xf));
+                scales = _mm_or_si128(_mm_slli_epi16(scales, 1), _mm_set1_epi8(1));
+                auto scales16 = _mm256_cvtepi8_epi16(scales);  // 0...7, 0...7
+#ifdef HAVE_FANCY_SIMD
+                __m256i scs[2] = { _mm256_shuffle_epi8(scales16, shuffles[0]), _mm256_shuffle_epi8(scales16, shuffles[1]) };
+                auto mask = (const __mmask32 *)&signs128;
+                for (int iy = 0; iy < nrc_y; ++iy) {
+                    auto y = _mm256_loadu_si256((const __m256i *)q8.y[iy][ibl].qs + ib);
+                    auto sumi1 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[0], _mm256_mask_sub_epi8(y, mask[0], _mm256_setzero_si256(), y)); // blocks: 0,0,0,0,  1,1,1,1, row 0
+                    auto sumi2 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[1], _mm256_mask_sub_epi8(y, mask[1], _mm256_setzero_si256(), y)); // blocks: 2,2,2,2,  3,3,3,3, row 1
+                    auto sumi3 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[2], _mm256_mask_sub_epi8(y, mask[2], _mm256_setzero_si256(), y)); // blocks: 4,4,4,4,  5,5,5,5, row 2
+                    auto sumi4 = _mm256_dpbusd_epi32(_mm256_setzero_si256(), qx[3], _mm256_mask_sub_epi8(y, mask[3], _mm256_setzero_si256(), y)); // blocks: 6,6,6,6,  7,7,7,7, row 3
+                    auto s12 = _mm256_packs_epi32(sumi1, sumi2);  // 0,0,0,0, 2,2,2,2,  1,1,1,1, 3,3,3,3
+                    auto s34 = _mm256_packs_epi32(sumi3, sumi4);  // 4,4,4,4, 6,6,6,6,  5,5,5,5, 7,7,7,7
+                    isum[2*iy+0] = _mm256_add_epi32(isum[2*iy+0], _mm256_madd_epi16(scs[0], s12));
+                    isum[2*iy+1] = _mm256_add_epi32(isum[2*iy+1], _mm256_madd_epi16(scs[1], s34));
+                }
+#else
+                auto signs = MM256_SET_M128I(signs128, signs128);
+                auto shuffle = sign_shuffle;
+                auto s1 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                shuffle = _mm256_add_epi8(shuffle, m4);
+                auto s2 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                shuffle = _mm256_add_epi8(shuffle, m4);
+                auto s3 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                shuffle = _mm256_add_epi8(shuffle, m4);
+                auto s4 = _mm256_or_si256(_mm256_cmpeq_epi8(_mm256_and_si256(_mm256_shuffle_epi8(signs, shuffle), smask), smask), _mm256_set1_epi8(1));
+                __m256i scs[4] = {
+                    _mm256_shuffle_epi8(scales16, shuffles[0]), _mm256_shuffle_epi8(scales16, shuffles[1]),
+                    _mm256_shuffle_epi8(scales16, shuffles[2]), _mm256_shuffle_epi8(scales16, shuffles[3]),
+                };
+                for (int iy = 0; iy < nrc_y; ++iy) {
+                    auto y = _mm256_loadu_si256((const __m256i *)q8.y[iy][ibl].qs + ib);
+                    if constexpr (nrc_y == 1) {
+                        isum[0] = _mm256_add_epi32(isum[0], _mm256_madd_epi16(scs[0], _mm256_maddubs_epi16(qx[0], _mm256_sign_epi8(y, s1))));
+                        isum[1] = _mm256_add_epi32(isum[1], _mm256_madd_epi16(scs[1], _mm256_maddubs_epi16(qx[1], _mm256_sign_epi8(y, s2))));
+                        isum[2] = _mm256_add_epi32(isum[2], _mm256_madd_epi16(scs[2], _mm256_maddubs_epi16(qx[2], _mm256_sign_epi8(y, s3))));
+                        isum[3] = _mm256_add_epi32(isum[3], _mm256_madd_epi16(scs[3], _mm256_maddubs_epi16(qx[3], _mm256_sign_epi8(y, s4))));
+                    } else {
+                        auto sumi1 = _mm256_madd_epi16(scs[0], _mm256_maddubs_epi16(qx[0], _mm256_sign_epi8(y, s1))); // blocks 4x0, 4x1, row 0
+                        auto sumi2 = _mm256_madd_epi16(scs[1], _mm256_maddubs_epi16(qx[1], _mm256_sign_epi8(y, s2))); // blocks 4x2, 4x3, row 1
+                        auto sumi3 = _mm256_madd_epi16(scs[2], _mm256_maddubs_epi16(qx[2], _mm256_sign_epi8(y, s3))); // blocks 4x4, 4x5, row 2
+                        auto sumi4 = _mm256_madd_epi16(scs[3], _mm256_maddubs_epi16(qx[3], _mm256_sign_epi8(y, s4))); // blocks 4x6, 4x7, row 3
+                        auto s12 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi1, sumi2), _mm256_unpackhi_epi32(sumi1, sumi2)); // 0,1, 0,1, 0,1, 0,1
+                        auto s34 = _mm256_add_epi32(_mm256_unpacklo_epi32(sumi3, sumi4), _mm256_unpackhi_epi32(sumi3, sumi4)); // 2,3, 2,3, 2,3, 2,3
+                        auto sumi = _mm256_add_epi32(_mm256_unpacklo_epi64(s12, s34), _mm256_unpackhi_epi64(s12, s34)); // 0,1,2,3, 0,1,2,3
+                        isum[iy] = _mm256_add_epi32(isum[iy], sumi);
+                    }
+                }
+#endif
+            }
+            for (int iy = 0; iy < nrc_y; ++iy) {
+#ifdef HAVE_FANCY_SIMD
+                auto sumi = _mm256_hadd_epi32(isum[2*iy+0], isum[2*iy+1]);
+                acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]);
+                isum[2*iy+0] = isum[2*iy+1] = _mm256_setzero_si256();
+#else
+                if constexpr (nrc_y == 1) {
+                    auto s12 = _mm256_add_epi32(_mm256_unpacklo_epi32(isum[0], isum[1]), _mm256_unpackhi_epi32(isum[0], isum[1]));
+                    auto s34 = _mm256_add_epi32(_mm256_unpacklo_epi32(isum[2], isum[3]), _mm256_unpackhi_epi32(isum[2], isum[3]));
+                    auto sumi = _mm256_add_epi32(_mm256_unpacklo_epi64(s12, s34), _mm256_unpackhi_epi64(s12, s34));
+                    acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]);
+                    isum[0] = isum[1] = isum[2] = isum[3] = _mm256_setzero_si256();
+                } else {
+                    acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(isum[iy]), acc[iy]);
+                    isum[iy] = _mm256_setzero_si256();
+                }
+#endif
+            }
+        }
+        for (int iy = 0; iy < nrc_y; ++iy) {
+            auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));
+            info.store(ix, iy, _mm_mul_ps(_mm_set1_ps(0.125f), sum));
+            acc[iy] = _mm256_setzero_ps();
+        }
+    }
+}
+
+template <int nrc_y>
 static void mul_mat_iq3_xxs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     GGML_ASSERT(nrc_x%4 == 0);
     Q8<nrc_y, block_q8_K> q8(info);
@@ -6938,6 +7060,18 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& mm, int Ny) {
             mm.funcs[7] = mul_mat_iq2_xs_r4_q8_k<8>;
             expected_typeB = GGML_TYPE_Q8_K;
             break;
+        case GGML_TYPE_IQ2_S_R4:
+            assert (ne00 % QK_K == 0);
+            mm.funcs[0] = mul_mat_iq2_s_r4_q8_k<1>;
+            mm.funcs[1] = mul_mat_iq2_s_r4_q8_k<2>;
+            mm.funcs[2] = mul_mat_iq2_s_r4_q8_k<3>;
+            mm.funcs[3] = mul_mat_iq2_s_r4_q8_k<4>;
+            mm.funcs[4] = mul_mat_iq2_s_r4_q8_k<5>;
+            mm.funcs[5] = mul_mat_iq2_s_r4_q8_k<6>;
+            mm.funcs[6] = mul_mat_iq2_s_r4_q8_k<7>;
+            mm.funcs[7] = mul_mat_iq2_s_r4_q8_k<8>;
+            expected_typeB = GGML_TYPE_Q8_K;
+            break;
         case GGML_TYPE_IQ3_XXS_R4:
             assert (ne00 % QK_K == 0);
             mm.funcs[0] = mul_mat_iq3_xxs_r4_q8_k<1>;
@@ -9939,6 +10073,72 @@ static void mul_mat_iq2_xs_r4_q8_k(int n, const void * vx, size_t bx, const Data
 }
 
 template <int nrc_y>
+static void mul_mat_iq2_s_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
+    GGML_ASSERT(nrc_x%4 == 0);
+    Q8<nrc_y, block_q8_K> q8(info);
+    int nbl = n / QK_K;
+    float32x4_t acc[nrc_y] = {};
+    int32x4_t   isum[2*nrc_y] = {};
+    int8x16_t   qx[8];
+    uint16x8x4_t scales16;
+    SignHelper  sh;
+    for (int ix = 0; ix < nrc_x; ix += 4) {
+        auto iq2 = (const block_iq2_s_r4 *)((const char *)vx + (ix+0)*bx);
+        for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
+            auto d4 = vcvt_f32_f16(vld1_f16((const float16_t *)iq2[ibl].d));
+            auto qs = iq2[ibl].qs;
+            auto qh = iq2[ibl].qh;
+            for (int is = 0; is < 2; ++is) {
+                auto scale_bits = vld1q_u8(iq2[ibl].scales + 16*is);
+                auto scales1 = vandq_u8(scale_bits, vdupq_n_u8(0xf));
+                auto scales2 = vshrq_n_u8(scale_bits, 4);
+                scales1 = vorrq_u8(vshlq_n_u8(scales1, 1), vdupq_n_u8(1));
+                scales2 = vorrq_u8(vshlq_n_u8(scales2, 1), vdupq_n_u8(1));
+                auto s1 = vzip1q_u8(scales1, scales2);
+                auto s2 = vzip2q_u8(scales1, scales2);
+                scales16.val[0] = vmovl_u8(vget_low_u8 (s1));
+                scales16.val[1] = vmovl_u8(vget_high_u8(s1));
+                scales16.val[2] = vmovl_u8(vget_low_u8 (s2));
+                scales16.val[3] = vmovl_u8(vget_high_u8(s2));
+                for (int ib = 0; ib < QK_K/64; ++ib) {
+                    auto signs128 = vld1q_u8(iq2[ibl].signs + 64*is + 16*ib);
+                    sh.init();
+                    for (int i = 0; i < 4; ++i) {
+                        qx[2*i+0] = vreinterpretq_s8_u64(uint64x2_t{iq2s_grid[qs[4*i+0] | ((qh[i] << 8) & 0x300)], iq2s_grid[qs[4*i+1] | ((qh[i] << 6) & 0x300)]});
+                        sh.apply_signs_1((uint8x16_t *)qx+2*i+0, signs128);
+                        qx[2*i+1] = vreinterpretq_s8_u64(uint64x2_t{iq2s_grid[qs[4*i+2] | ((qh[i] << 4) & 0x300)], iq2s_grid[qs[4*i+3] | ((qh[i] << 2) & 0x300)]});
+                        sh.apply_signs_1((uint8x16_t *)qx+2*i+1, signs128);
+                    }
+                    qs += 16; qh += 4;
+                    auto s32_1 = vmovl_u16(vget_low_u16 (scales16.val[ib]));
+                    auto s32_2 = vmovl_u16(vget_high_u16(scales16.val[ib]));
+                    for (int iy = 0; iy < nrc_y; ++iy) {
+                        auto y = vld1q_s8_x2(q8.y[iy][ibl].qs + 128*is + 32*ib);
+                        auto sumi1 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[0], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[1], y.val[1]));
+                        auto sumi2 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[2], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[3], y.val[1]));
+                        auto sumi3 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[4], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[5], y.val[1]));
+                        auto sumi4 = vpaddq_s32(ggml_vdotq_s32(vdupq_n_s32(0), qx[6], y.val[0]), ggml_vdotq_s32(vdupq_n_s32(0), qx[7], y.val[1]));
+                        auto sumi12 = vpaddq_s32(sumi1, sumi2); // blocks 0,1,2,3 in rows 0,1
+                        auto sumi34 = vpaddq_s32(sumi3, sumi4); // blocks 4,5,6,7 in rows 2,3
+                        isum[2*iy+0] = vmlaq_s32(isum[2*iy+0], s32_1, sumi12);
+                        isum[2*iy+1] = vmlaq_s32(isum[2*iy+1], s32_2, sumi34);
+                    }
+                }
+            }
+            for (int iy = 0; iy < nrc_y; ++iy) {
+                auto sumi = vpaddq_s32(isum[2*iy+0], isum[2*iy+1]);
+                acc[iy] = vfmaq_f32(acc[iy], vmulq_f32(d4, vdupq_n_f32(q8.scale(iy, ibl))), vcvtq_f32_s32(sumi));
+                isum[2*iy] = isum[2*iy+1] = vdupq_n_s32(0);
+            }
+        }
+        for (int iy = 0; iy < nrc_y; ++iy) {
+            info.store(ix, iy, vmulq_f32(vdupq_n_f32(0.125f), acc[iy]));
+            acc[iy] = vdupq_n_f32(0.f);
+        }
+    }
+}
+
+template <int nrc_y>
 static void mul_mat_iq3_xxs_r4_q8_k(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
     GGML_ASSERT(nrc_x%4 == 0);
     Q8<nrc_y, block_q8_K> q8(info);
@@ -11293,6 +11493,10 @@ bool MulMat::prepare(int typeA, int typeB, int ne00, MulMat& m, int /*Ny*/) {
             SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq2_xs_r4_q8_k);
             expected_Btype = GGML_TYPE_Q8_K;
             break;
+        case GGML_TYPE_IQ2_S_R4:
+            SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq2_s_r4_q8_k);
+            expected_Btype = GGML_TYPE_Q8_K;
+            break;
         case GGML_TYPE_IQ3_XXS_R4:
             SET_MUL_MAT_FUNCTIONS(m, mul_mat_iq3_xxs_r4_q8_k);
             expected_Btype = GGML_TYPE_Q8_K;
diff --git a/ggml/src/iqk/iqk_quantize.cpp b/ggml/src/iqk/iqk_quantize.cpp
index e369a2f0..6683ef14 100644
--- a/ggml/src/iqk/iqk_quantize.cpp
+++ b/ggml/src/iqk/iqk_quantize.cpp
@@ -5507,6 +5507,93 @@ void vec_dot_iq2_xs_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t
 }
 
 //
+// ========================================= iq2_s_r4
+//
+
+void quantize_row_iq2_s_r4_ref(const float * x, block_iq2_s_r4 * y, int64_t k) {
+    quantize_iq2_s_r4(x, (void *)y, 4, k/4, nullptr);
+}
+
+void quantize_row_iq2_s_r4(const float * x, void * y, int64_t k) {
+    quantize_iq2_s_r4(x, y, 4, k/4, nullptr);
+}
+
+static void repack_iq2_s(int nrows, int n_per_row, const block_iq2_s * x, block_iq2_s_r4 * y) {
+    GGML_ASSERT(nrows%4 == 0);
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int nblock = n_per_row/QK_K;
+    const block_iq2_s * x4[4];
+    for (int row = 0; row < nrows; row += 4) {
+        for (int k = 0; k < 4; ++k) x4[k] = x + nblock*k;
+        for (int ibl = 0; ibl < nblock; ++ibl) {
+            for (int k = 0; k < 4; ++k) {
+                auto signs = x4[k][ibl].qs + QK_K/8;
+                y[ibl].d[k] = x4[k][ibl].d;
+                for (int ib = 0; ib < QK_K/32; ++ib) {
+                    y[ibl].scales[4*ib+k] = x4[k][ibl].scales[ib];
+                    for (int i = 0; i < 4; ++i) {
+                        y[ibl].qs[16*ib+4*k+i] = x4[k][ibl].qs[4*ib+i];
+                        y[ibl].signs[16*ib+4*k+i] = signs[4*ib+i];
+                    }
+                    y[ibl].qh[4*ib+k] = x4[k][ibl].qh[ib];
+                }
+            }
+        }
+        x += 4*nblock;
+        y += nblock;
+    }
+}
+
+size_t quantize_iq2_s_r4(const float * src, void * dst, int64_t nrows, int64_t n_per_row, const float * imatrix) {
+    GGML_ASSERT(nrows%4 == 0);
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    char * qcur = (char *)dst;
+    auto row_size = ggml_row_size(GGML_TYPE_IQ2_S, n_per_row);
+    std::vector<char> qtmp(4*row_size);
+    for (int row = 0; row < nrows; row += 4) {
+        quantize_iq2_s(src, (void *)qtmp.data(), 4, n_per_row, imatrix);
+        repack_iq2_s(4, n_per_row, (const block_iq2_s *)qtmp.data(), (block_iq2_s_r4 *)qcur);
+        qcur += 4*row_size;
+        src += 4*n_per_row;
+    }
+    return nrows*row_size;
+}
+
+void dequantize_row_iq2_s_r4(const block_iq2_s_r4 * x, float * y, int64_t k) {
+    auto n_per_row = k/4;
+    float * y4[4] = {y, y + n_per_row, y + 2*n_per_row, y + 3*n_per_row};
+    int nblock = n_per_row/QK_K;
+    for (int ibl = 0; ibl < nblock; ++ibl) {
+        for (int k = 0; k < 4; ++k) {
+            const float d = 0.125f*GGML_FP16_TO_FP32(x[ibl].d[k]);
+            for (int ib = 0; ib < QK_K/32; ++ib) {
+                float dl1 = d * (2*(x[ibl].scales[4*ib+k] & 0xf) + 1);
+                float dl2 = d * (2*(x[ibl].scales[4*ib+k] >>  4) + 1);
+                for (int i = 0; i < 4; ++i) {
+                    auto val = (const int8_t *)(iq2s_grid + (x[ibl].qs[16*ib+4*k+i] | ((x[ibl].qh[4*ib+k] << (8 - 2*i)) & 0x300)));
+                    auto signs = x[ibl].signs[16*ib+4*k+i];
+                    float dl = i < 2 ? dl1 : dl2;
+                    for (int j = 0; j < 8; ++j) y4[k][QK_K*ibl+32*ib+8*i+j] = dl * val[j] * (signs & (1 << j) ? -1 : 1);
+                }
+            }
+        }
+    }
+}
+
+void vec_dot_iq2_s_r4_q8_k(int n, float * s, size_t bs, const void * vx, size_t bx, const void * vy, size_t by, int nrc) {
+#if GGML_USE_IQK_MULMAT
+    if (iqk_mul_mat(1, 1, n, GGML_TYPE_IQ2_S_R4, vx, 0, GGML_TYPE_Q8_K, vy, 0, s, 0, 0, 1)) {
+        return;
+    }
+#endif
+    GGML_ASSERT(n%QK4_NL == 0);
+    GGML_ASSERT(nrc == 1);
+    GGML_UNUSED(bs);
+    GGML_UNUSED(bx);
+    GGML_UNUSED(by);
+}
+
+//
 // ========================================= iq3_xxs_r4
 //
 
diff --git a/ggml/src/iqk/iqk_quantize.h b/ggml/src/iqk/iqk_quantize.h
index 7b183956..f62f055a 100644
--- a/ggml/src/iqk/iqk_quantize.h
+++ b/ggml/src/iqk/iqk_quantize.h
@@ -181,6 +181,12 @@ size_t quantize_iq2_xs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT
 void   dequantize_row_iq2_xs_r4(const block_iq2_xs_r4  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 void   vec_dot_iq2_xs_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
+void   quantize_row_iq2_s_r4_ref(const float * GGML_RESTRICT x, block_iq2_s_r4  * GGML_RESTRICT y, int64_t k);
+void   quantize_row_iq2_s_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+size_t quantize_iq2_s_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+void   dequantize_row_iq2_s_r4(const block_iq2_s_r4  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void   vec_dot_iq2_s_r4_q8_k(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
 void   quantize_row_iq3_xxs_r4_ref(const float * GGML_RESTRICT x, block_iq3_xxs_r4  * GGML_RESTRICT y, int64_t k);
 void   quantize_row_iq3_xxs_r4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 size_t quantize_iq3_xxs_r4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);